Grok 5 will be the decisive model that achieves artificial general intelligence and outperforms all previous models in technology history.
Article index:
- Musk’s announcement on X and the promise to reach general intelligence
- Grok 4.8 specifications and software stack update
- Comparison of Grok 4.7 with Opus and Astra models
- Delay in rolling out Grok 4.7 and reasons for setbacks in reinforcement learning
- The race of successive releases at Anthropic, Google, and OpenAI
- Contradiction of promises with de-escalation calls and chip stock drop
- Frequently asked questions
Musk’s announcement on X and the promise to reach general intelligence
Billionaire Elon Musk announced on Monday that the upcoming future version “Grok 5,” developed by his startup xAI, will be the historic software model that succeeds in achieving “artificial general intelligence,” setting a highly ambitious and bold roadmap at a time when the company is still lagging behind in meeting the release date of its previous version, “Grok 4.7.”
Musk’s statements came through a series of consecutive posts on the X platform on September 14, where he revealed that the “Grok 4.8” model, which features 2.5 trillion compute parameters and is built on an updated C++ software stack, is nearing the completion of its initial training phase in just one week before entering the reinforcement learning stage. When asked how close this model is to general intelligence, he replied concisely and confidently: “That achievement will belong to Grok 5.”
Grok 4.8 specifications and software stack update
xAI is betting on completely rebuilding its software architecture to bypass the computational limitations of previous models. The transition to an architecture built natively in C++ allows for faster token invocation and improved direct communication among hundreds of thousands of GPUs inside the company’s giant data centers.
The model’s scale of 2.5 trillion parameters reflects Musk’s desire to exceed current model sizes and double the depth of mathematical and logical reasoning, despite the massive challenges associated with high electric power consumption and the need for tremendous cooling for continuous training servers.
Comparison of Grok 4.7 with Opus and Astra models
Musk described the unreleased “Grok 4.7” model as “roughly equivalent in performance to Anthropic’s recently launched Claude Opus 5.0 model, not Opus 5.1,” adding realistically that his model is “better in some ways and weaker in others,” while acknowledging that multimodal processing efficiency such as images and video still requires further work and improvement.
Musk predicted that the subsequent “Grok 4.9” model would rise to the competitive top tier to match OpenAI’s “GPT-6 Astra” and Anthropic’s “Fable” series, describing Grok 5 as “potentially better than any model ever,” and concluding his remarks with a cautious phrase: “We will see what happens,” without providing any approved benchmark test results, a binding timeline, or a precise scientific definition of what reaching general intelligence means.
Delay in rolling out Grok 4.7 and reasons for setbacks in reinforcement learning
This boasting about the future of Grok 5 comes against the backdrop of an embarrassing and confusing operational background for the company. Grok 4.7 was scheduled to launch around September 12 but has not seen the light of day yet. Musk justified this delay on September 11 by stating that the model needs “a few extra days to bake in the training oven,” revealing that the engineering team may have penalized the model too harshly for answer length during reinforcement learning training.
This programming bug led the model to lean toward early surrender when facing complex problems and laziness in reviewing and auditing its own solution steps, leaving the latest commercially available model for users as “Grok 4.6,” which launched on August 12 at two dollars per million input tokens via the developer interface.
The race of successive releases at Anthropic, Google, and OpenAI
xAI’s setback looks striking given the intense vitality shown by competitors in Silicon Valley. Anthropic rolled out the “Claude Fable 5.1” model on September 1, Google DeepMind published the technical card for the “Gemini 3.8 Flash” model on September 2, and OpenAI presented the safety preview for the “Astra” model on September 3, making the field witness the launch of three frontier models in three consecutive days while xAI is still struggling to fine-tune its delayed model.
Despite the company possessing the massive “Colossus” computer cluster in Memphis housing 200,000 advanced Nvidia GPUs, industry analysts emphasize that “compute power gives you a ticket to enter the race, but by itself it cannot print the commercial launch date” without mastering training algorithms.
Contradiction of promises with de-escalation calls and chip stock drop
The timing of Musk’s announcement sparked widespread surprise, as it came on the same day he publicly endorsed Dario Amodei and Sam Altman’s calls to slow down the pace of artificial intelligence development for national security reasons, leading observers to raise questions about how to reconcile demanding de-escalation while chasing general intelligence at the same time.
These contradictions reflected negatively on financial markets; Nvidia shares fell by 3.46 percent and AMD shares dropped by 6.24 percent following investor concerns over a slowdown in chip purchases, leaving Musk’s promises pending tangible practical proof.
Frequently asked questions
Question: What is the exciting announcement made by Elon Musk regarding the Grok 5 model?
Answer: Musk announced that the Grok 5 model will be the software system that officially succeeds in achieving artificial general intelligence and outperforming globally.
Question: Why was the rollout of the Grok 4.7 model delayed from its scheduled September date?
Answer: Musk explained that the model underwent strict penalties on answer length in reinforcement learning, making it surrender early and requiring further tuning.
Question: What hardware compute power does xAI rely on in training?
Answer: The company relies on the giant Colossus cluster in Memphis, which houses 200,000 advanced Nvidia H100 graphics processing units.
